Optimising cohort data in Europe

balance must be found between accuracy and validity among the different variables in each study in order to construct the new common variable. Most prominently, pooled analyses adjust for the centre using either fixed or random effects on the outcome. Fixed effect models are preferred when having large sample sizes per study and the different cohorts provide samples regarding the same target population. On the other hand, random effect models are preferred when there are some studies with a small number of subjects and when there is much difference between the studied populations. The two‐stage approach first analyses the individual participant data at each centre separately and then combines the results across studies. For this second stage, this approach is also called “aggregated data meta-analysis”. The aim of the first stage is to produce estimations of the statistics of interest with its variance in each study. In the second stage, the coefficient estimates obtained in the first stage are combined. Again, they can be considered either fixed or random across studies. When the effect of interest is considered fixed, the estimates from the previous analyses are considered as observations of normal variables, and the most common method to estimate the single effect across studies is to consider the average of the individual effects weighted by the inverse of their variances. Alternatively, when the effect of interest is considered random, the estimate of the effect across centres is again a weighted average of these individual effects but, in this case, the weights are usually given by the inverse of the sum between variances at each centre and the between study-variance, a measure of the between-study heterogeneity. The main advantage of the two-stage meta-analysis is that it accounts for between-study heterogeneity. The one-staged pooled analysis can also account for it when necessary, but its main advantage is its higher statistical power, which derives from the reduction of parameters to estimate and the potential to better modelling the combined data. Other important differences include: y y The two-stage approach assumes that the estimates distribute normally with known variances while the one-stage approach actually models the individual participant data and does not rely on this assumption. y y The two-stage approach is problematic for studies on rare outcomes, with low number of participants or unbalanced subgroups. Different weighting averages can be used to correct for these issues. y y The two-stage approach accounts for the clustering of patients by trial. The one stage approach needs to explicitly account for the clustering by stratifying the analysis by trial or assuming a random-effect for the intercept across trials. y y The one-stage approach may overlook that the effect of adjustment for covariates could be different between studies. This is automatically done in the two-stage approach but needs to be explicitly defined in one-stage analyses.